Create a release¶
IceNet-MP uses CalVer of the format YYYY.MM (e.g. 2026.07) for version numbers. Releases are made at least once per milestone, or sooner if there have been significant changes or improvements to the model.
1. Prepare the release¶
Create a new branch containing the changes to be included in the release. Increment the version in pyproject.toml and any other metadata that differs for the new release.
Open a pull request and merge it to main.
2. Publish the release on GitHub¶
On the main repo page, go to Releases -> Draft a new release.
- Click Choose a tag and create a new tag named the version number.
- Choose the commit to target the release at (the one you just merged to
main). - Fill in the descriptive fields and publish the release.
3. Run the model comparison suite¶
Once the release has been published, train the default set of models for a 24-hour run on Isambard, using the tagged release version. This should use the full dataset, with separate runs for the northern hemisphere and the southern hemisphere, and all other defaults. The models tested should be the basic model suite, i.e.:
- persistence (or similar climatology baseline)
- the four best performing models at release time
For each training run, set the W&B run name to the version and the model type (e.g. 2026-07-persistence) via the local config:
Evaluate each trained model, again using the version-based name for the evaluation run, and upload summary evaluation plots/metrics to GitHub as release attachments.
Finally, copy the trained/evaluated run folders to Sharepoint as a backup.